Posted on 06/16/2004 8:17:00 AM PDT by Darth Reagan
I'm not using the term normalize precisely, so I should have been more clear. Here's a precise definition: "A collection of numeric data is normalized by subtracting the minimum value from all values and dividing by the range of the data. This yields data with a similarly shaped histogram but with all values between 0 and 1." This is not at all what I'm doing, but I'm following the spirit of normailization, which is to allow you do compare different sets of data.
The basic idea is you have different sets of data and want to compare them so that you have an apples to apples comparison. For example, Mason-Dixon showing Bush ahead by 2 in Pennsylvania would be the same as Rasmussen showing Kerry ahead by 2.
To normalize (again, I'm not using this precisely -- there may be a better word -- but basically adjust the data of the polls so they are comporable) the data, what I did was pick a poll as a base line (say SUSA) and then compare how the different polls compare to that state by state, and then take the average of that. For example, we'll look at SUSA and Rasmussen for a couple of states:
1) FL +7 0
2) AR +2 0
3) MI -4 -6
4) NC +11 +4
5) CA -1 -8
The first number is SUSA, the second RAS. In this particular sample of 5 states, on the average RAS is 5 points more favoreable towards Kerry than SUSA is. If you did this process for all the states, you would get a result of 3.8, meaning that if you add 3.8 to a Rasmussen result for Bush (or add 3.8 to Kerry for SUSA if you're comparing in the other direction), then the two polls will be comparable.
In the tracking I've done, SUSA has been most in line with what I would have expected. However, they do have a lot of polls that are off one direction or the other (e.g. the CA poll showing Bush behind by 1), but these differences offset so that on the average there is no bias towards Kerry. Other state polls that haven't shown a bias are Quinnipiac, Mason-Dixon, and the L.A. Times (!). Polls which have are Rasmussen, ARG, many of the University and newspaper polls, and the worst offender of all, Zogby.
Now in terms of "normailization" one can pick whatever poll one likes as the standard. The most accurate ones last time were Mason-Dixon and SUSA, and they appear to me (along with Quinnipiac) to be the most accurate this time around too.
The advantage of "normalizing" the data is you get a consistent picture, rather than a snap-shot which makes it looks like there's been movement when there really hasn't been. (the "movement" was just who happened to have done the last poll). For example, if SUSA and Rasmussen one day apart do a poll in PA and they show Bush ahead by 2 and Kerry ahead by 2 respectively it will look like Kerry has picked up 4 points and is now in the lead in Ohio (or the reverse if the SUSA one was done later) when in reality the two polls are saying the same thing.
Hopefully that clarifies things.
You may also want to combine error estimates. The linked site gives methods for doing this.
Thanks. You are right that what I am doing is close to this, but there is an extra step here that I will try to incorporate.
You know, I love your style and, as you know, have even identified your stuff by your headlines.
But what is your record as a predictor? Exactly?
Dan
Applying your own weights to reflect your judgment of the polls reliability is also valid. You do have to adjust the variances though.
This is what all the pollsters do. Zogby admits it. He adjusts weights to reflect what he perceives to be a more accurate voting pattern than a raw count would.
There are several articles (about 10 year ago) in the Journal of the Royal Statistical Society about problems with sampling for elections.
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.